NVIDIA's Auto Ontology Builds SQL Context Layers, Won't Accept Outside Code

NVIDIA publishes Auto Ontology, an Apache-2.0 Python framework that builds semantic, graph-based context layers over enterprise metadata so agents can generate SQL against certified definitions and return provenance. It stores the catalog in Postgres with pgvector, defaults to NVIDIA NIM models, and queries Databricks, Snowflake and DuckDB sources. The repository states 462 stars and that outside contributions are not accepted.

• The repository says connection credentials sit in plaintext in Postgres unless all four Vault variables are set.
• The project stores API tokens only as a SHA-256 hash, which it shows once at creation.
• Agent harnesses such as Cursor and Claude Desktop reach data through the MCP server in the mcp directory.